This page explains how the Proven Agent Report Card scores B2B SaaS vendors, what each probe measures, how scores are weighted, how often they're updated, and how to request a correction or rescore.
The Agent Report Card answers one question: how well can an AI agent discover, understand, and use this vendor's product without human help?
Agents need different things than human users. They can't fill out a demo request form. They can't navigate a JavaScript-heavy pricing page. They need machine-readable APIs, structured data, and protocol support. The score measures whether those things exist — not whether the product itself is good.
A high score means an agent can autonomously discover the vendor, read their pricing, call their API, and integrate with their product. A low score means the vendor is effectively invisible to agent-driven workflows — even if the product is excellent.
Can agents reach and read the vendor's public web presence without being blocked? This dimension checks three signals:
| Probe | Points | What we check |
|---|---|---|
llms.txt | 5 | A structured guide for AI agents at /llms.txt — the emerging convention for telling agents what's important on a site. |
robots.txt | 4 | Whether robots.txt explicitly allows common AI bots: GPTBot, ClaudeBot, PerplexityBot, Google-Extended. A missing or blocking robots.txt fails. |
Sitemap | 3 | A discoverable sitemap.xml at standard paths, so agents can find the vendor's full URL set. |
Can agents call the vendor's API without requiring OAuth flows or human-mediated access?
| Probe | Points | What we check |
|---|---|---|
OpenAPI spec | 12 | A public machine-readable API definition at standard paths (/openapi.json, /swagger.json, etc.). Quality bonus: defined auth, examples, error schemas, 5+ endpoints. |
Docs site | 8 | A separate developer docs subdomain (docs.X or developers.X) where agents can find the API reference. |
Can agents understand what the product costs and what each plan includes?
| Probe | Points | What we check |
|---|---|---|
Pricing page | 10 | A pricing page structured enough that agents can extract plans, prices, and limits without guessing. We run an AI-assisted readability score on the page content — partial credit for pages that exist but are hard to parse. |
Does the vendor's site give agents the context they need to understand who they are and what they do?
| Probe | Points | What we check |
|---|---|---|
Schema.org | 8 | JSON-LD structured data on the homepage. We look for SoftwareApplication, Organization, Product, or similar types that help agents understand the vendor's identity and offering. |
Is the vendor plugged into the protocol and tooling layer that agents actually use to take action?
| Probe | Points | What we check |
|---|---|---|
MCP server | 8 | Listed in community MCP server directories (modelcontextprotocol/servers, punkpeye/awesome-mcp-servers). |
Official SDK | 6 | An official SDK published on GitHub in the vendor's namesake org with active maintenance. |
LangChain / LlamaIndex | 6 | First-class integration in dominant agent-framework ecosystems. |
Agent-friendly auth | 5 | Whether the API auth model supports headless agents — API keys with scoped permissions beat OAuth-only flows. |
Webhook support | 5 | Documented webhooks with payload schemas and signature verification, so agents can react to events without polling. |
Agent skills repo | 5 | An official GitHub repo publishing reusable skills for Claude Code and agent-skills-compatible tools. |
Context7 listing | 3 | Listed in Context7's developer-documentation index, making docs available to AI assistants via context7.com. |
Raw probe points are summed and normalized to a 0–100 scale based on the maximum achievable points across all probes a vendor was eligible for. Vendors blocked by WAF or Cloudflare at the crawlability layer receive a partial score computed only over the dimensions that could be measured — the score is normalized to 100 based on that reduced point pool.
| Score range | Grade | What it means |
|---|---|---|
| 70–100 | A / B | Well-positioned for agent-driven workflows. Agents can likely discover, evaluate, and use this vendor autonomously. |
| 40–69 | C / D | Partial readiness. Some gaps that would prevent full autonomous use — typically missing API spec or agent ecosystem signals. |
| 0–39 | F | Significant gaps. Agents are unlikely to discover or use this vendor without human-assisted steps. |
Scores are refreshed periodically as vendors ship changes. Each vendor detail page shows the date the score was last computed. Vendors who have made improvements can request a rescore — see below.
We do not re-score all vendors simultaneously. Scores reflect the state of the vendor's public presence at the time of the most recent probe run, which may be weeks or months old for less frequently updated vendors.
All probes run against publicly accessible URLs — no login, no vendor cooperation required. This has two important implications:
The score measures what an AI agent would encounter in the wild — not what exists internally. Vendors who want their full readiness recognized should publish and list their integrations publicly.
If your company is already listed and you've made improvements since the last score, use the Request a ranking button on the marketplace — rescores typically run within a few business days.
If you believe a specific probe result is inaccurate (e.g., we missed an OpenAPI spec that is publicly accessible), email hello@proven.email with the probe name, your domain, and the URL that should be detected. We'll investigate and correct within 5 business days if the evidence supports it.
Proven operates a commercial vendor marketplace used by VC-backed startups and their investors. Some vendors listed in the Agent Report Card may also appear in Proven's commercial marketplace or have existing relationships with Proven.
Scores are computed entirely by automated probes against public signals — no vendor can pay to improve their score, and Proven staff do not manually adjust scores based on commercial relationships. The methodology and probe logic are applied identically across all vendors.
If you believe a score has been manipulated in bad faith, contact hello@proven.email.